What Is an Adapter Layer in AI?

Adapter Layer in AI Explainer

An adapter layer is a compact trainable module added between or inside layers of a pretrained neural network. During adaptation, the original network is usually frozen while the adapter learns how to modify intermediate representations for a task.

Many adapters use a bottleneck: they project a hidden state into a smaller dimension, apply a transformation, and project it back. A residual connection lets the module add a controlled change rather than replace the base representation.

Adapters reduce the amount of task-specific storage and can make one base model serve many domains. Modules may also be composed, switched, or combined, although compatibility depends on architecture, training procedure, and the intended interaction.

Added modules can introduce inference overhead, conflict when combined, or underperform full fine-tuning. They also inherit limitations from the base model. Evaluation should cover individual tasks, module composition, latency, drift, and version compatibility.

An adapter layer is a small trainable module inserted into a pretrained network while most original model weights stay frozen.

Key terms include adapter layer, bottleneck, projection, residual connection, frozen model, module composition, transfer and latency.

The adapter transforms intermediate features while most of the pretrained network remains fixed.

A bottleneck limits the number of task-specific parameters and can preserve modularity.

Adapters may depend on a particular architecture, layer placement, and base-model version.

Combined modules require testing because their learned changes can interfere.

Explore more "Explainers"

Discover additional explainers across politics, science, business, technology, and other fields. Each explainer breaks down a complex idea into clear, everyday language—helping you better understand how major concepts, systems, and debates shape the world around us.